WORDING

Decision Arena: wording rules

The report generator and every page or post built from an arena run follow these rules. They exist because the sample is small and the cases were written, and tuned, by RISE's authors.

1. Agreement, not accuracy. Say "agreement with author-written expectations". Never "accuracy", "correct", or "right answers" for the 39 authored cases: the expectations are one author's taste, not ground truth.
2. No ranking inside the noise. When a paired bootstrap interval includes zero, is undefined, or rests on fewer than five shared cases (withinNoise: true from scripts/arena/calibration.mjs), do not order the deciders, name a winner, or use "better", "beats", or "leads". Say "no difference detectable at n = …".
3. Calibration v1 tests stated-odds mechanics, not taste. The known-probability controls check whether a decider puts the stated odds on the stated options. They say nothing about whether its reading choices are good.
4. Results are per model id, date and release. Every number carries the exact model id, the capture date and the RISE release (commit). Never generalise to a vendor or a model family, and say that frozen results age.
5. Never "calibrated". At most: "no miscalibration detectable at n = …" when the ECE sits below its simulated noise floor and the Brier interval is consistent with the stated odds. Report confidence and per-option probabilities separately; never merge them into one figure.
6. Show the interval with the number. Every Brier score, agreement rate and difference is printed with its 95% interval and its n. Reliability bins with fewer than 15 rows are marked as thin.
7. Never pool repeats. The runs of one decider ask the same cases, so they are not independent; an interval over all runs pooled is falsely narrow. State run 1 and show the other runs beside it as stability. Take agreement intervals and differences from the report's agreement block (node scripts/arena/arena.mjs report), never compute them on the page: explicit agreement carries a case-clustered bootstrap interval, contrast pairs a Wilson interval over run 1's pairs, and each difference a paired case-clustered interval with withinNoise.

---

Published from docs/decision-arena/WORDING.md. Edits made here are overwritten on the next push to main.